DISPUTED: no
Self-evolving Agents
Summary
Self-evolving agents are AI systems that improve their own performance over time without direct human retraining. A taxonomy by @atasteoff classifies these systems into three levels based on what evolves: artifacts (outputs), agent harness (prompts, memory, tools), or model parameters. The framework centers on three factors – Models (e.g., LLMs), Harness (loop designs, memory, tools), and Artifacts (outputs like code, papers, robot policies) – and examines how different approaches close the improvement loop using weak signals, environment feedback, or internal signals.
Key Points
- Self-evolving agents are distinguished by what evolves: artifacts (outputs), agent harness (prompts, memory, tools/skills), or model weights.
- Models + Harness form agents; agents produce artifacts. Evolution can target any of these components.
- Three major levels: Artifact Iterative Optimization, Agent Harness Self-improvement, and Model Learning without Gold Answers.
- Examples include AlphaEvolve (artifact optimization), Hermes Agent (skill creation), RSI Lab (recursive algorithm discovery), and DeepSeek-R1 (self-reward learning).
- The boundaries between levels are blurring; future systems will likely evolve all three components together.
- The ultimate value is measured by real-world improvements: faster kernels, better software, new scientific hypotheses, or physical-world actions.
Concepts
Three Factors:
- Models: Usually LLMs that provide the core intelligence.
- Harness: All components that turn a model into an agent – loop designs, memory systems, prompts, tools, skills, and orchestration. Agent = Model + Harness.
- Artifacts: Outputs produced by agents, such as algorithms, research papers, robot policies, or scientific discovery results.
Three Levels of Self-evolving Systems:
- Artifact Iterative Optimization: The agent repeatedly improves an output artifact (e.g., code, paper) against a human-set criterion. The harness and model remain fixed.